Alright, here’s the no-bullshit version from an old /adt/ anon:

Models

- novaAnimeXL_ilV90: decent with regional prompting, but easy to overconstrain and wash out if you cram too many attrs.
- Inpaint: use an actual inpaint model or ControlNet Inpaint. If editing a hand-drawn piece, match style (base+style ref), or you’ll fight seams forever.

LoRAs, tags, prompts
- ComfyUI: `<lora:Name:0.5>` in text doesn’t do shit unless you use parser nodes. Load LoRA via node; use trigger tag only in text.
- Synonyms mostly work; don’t be a slave to booru tags. But some tags nuke style: “young” often forces bean-mouth/round eyes. Know your model’s quirks.
- Character oddities (e.g., Ryuko’s pupils) come from data. Want normal? Negative it or override with a style LoRA.
- Adetailer: run on face, not just eyes. Eyes-only crops can shift colors. If it still derps, change its inpaint model and strip conflicting style from its sub-prompt. Bad LoRAs cause 90% of eye pain.

Inpaint/workflows
- Forge is snappier for quick inpaint. Krita plugin is only good if you build a clean Comfy custom workflow and expose minimal knobs. Otherwise, stick to your main UI.
- Big edits: use inpaint model + pose/depth ControlNet. Small fixes: rough it by hand, then inpaint to clean.